Overview Features Coding ApolloOS Performance Forum Downloads Products Order Contact

Welcome to the Apollo Forum

This forum is for people interested in the APOLLO CPU.
Please read the forum usage manual.
Please visit our Apollo-Discord Server for support.



All TopicsNewsPerformanceGamesDemosApolloVampireAROSWorkbenchATARIReleases
Performance and Benchmark Results!

V1200 Updatepage  1 2 

Gunnar von Boehn
(Apollo Team Member)
Posts 6207
29 May 2020 16:25


Andy Hearn wrote:

A Sonnet card loaded with 256 meg ram, a PPC 7410 CPU - G4 @450Mhz generates the following numbers in amiga SysSpeed

The problem of SysSpeed is that the executed "workload" the done code - on PPC and on 68K is not the same.

So you have you compare 2 different CPUs properly - if you NOT do the same work on them?

And even if you do the same work, how do you then count MIPS?

If you now for example see that for
this 68K instruction


ADDi.L #1,(a0)+

The PPC code needs comparable 3 instructions


LOAD (Rn),Rm
ADD #1, Rm
STORE Rm,(Rn)+

Its the same work done, but 3 times the number of instructions.
What does SysSpeed count here 1 or 3?

You can see comparing cross CPU architecture is complex topic,
I know this I have written CPU PowerPC benchmarks at IBM.

What is much easier to measure are real life workloads.
For example run BUSTEST and see how fast memspeed is.
It does not matter if the 68K needs halve the instruction to do the job - so not count instructions or Mips, count the result.
Count the MB/sec

So run your SONNET and look at BUSTEST scores to see "real life performance".

Or display a JPEG .. such numbers are much more real to measure.


Renee Cousins
(Apollo Team Member)
Posts 142
29 May 2020 21:49


Congrats to Igor.

Gunnar von Boehn wrote:
And even if you do the same work, how do you then count MIPS?

MIPS is a measure of relative performance using the Dhrystone 2.1 test from a VAX 11/780 which was considered a 1 MIPS machine. It is not the literal number of opcodes executed per second and should never be interpreted as meaning that.

Using the real Dhrystone test is the only way to compare across legacy architectures and neither SysInfo nor SysSpeed measure real Dhrystone MIPS -- they're basically made up numbers that BARELY have any use measuring relative performance within the same CPU family.

Gunnar von Boehn wrote:
What is much easier to measure are real life workloads. For example run BUSTEST and see how fast memspeed is.

Sequential memory bandwidth is just one measure and it's only helpful if you understand other factors like code density and penalties for cache misses.

I'm not saying MIPS is great either -- it's subject to optimization abuse -- but until they port CoreMark to 68000, we're a little short on cross-platform measurement tools.

It's important to not accept chip vendors at their word either -- grab a copy of the Dhrystone 2.1 program and run it yourself.


Vojin Vidanovic
(Needs Verification)
Posts 1916/ 1
29 May 2020 22:45


Aminet hosts quite old Drystone 1.1
 
  Short: Dhrystone Benchmark, all Amigas
  Uploader: dcc dcs ed ac uk
  Architecture: m68k-amigaos
  Date: 1993-04-20
  Download: EXTERNAL LINK - View contents
  Readme: EXTERNAL LINK  Downloads: 522
 
  This is the Dhrystone benchmark compiled to be compatible with all Amigas.
  It may be useful if your machine will not run SysInfo. Note that since this is highly optimised, it gives higher results than SysInfo's built-in Dhrystone.
 
  There is some UNIX ported one 1995 which might include 2.1
  EXTERNAL LINK 


Markus B

Posts 209
30 May 2020 04:46


I assume the chipmem used on the V1200 setup is the one from the mainboard, right?


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
30 May 2020 07:19


Vojin Vidanovic wrote:

Aminet hosts quite old Drystone 1.1

 
But this tool does not help us here.
As we discussed Dhrystone depends a lot on compiler optimization, and on the linked string libraries. This is why companies have their own team to set these libs and compile options correctly up for compiles.
This old compile on Aminet is not at all useful to measure the V1200 update and not very useful for comparing CPUs.
 
 
The topic of this thread is the V1200 bustuning.
The tool to measure this this most easy is BUSTEST.
You can also use SYSINFO and look at this CHIPSPEED score.
 
And of course real life programs will also show improvements.



Gunnar von Boehn
(Apollo Team Member)
Posts 6207
30 May 2020 07:22


Renee Cousins wrote:

  Sequential memory bandwidth is just one measure and it's only helpful if you understand other factors like code density and penalties for cache misses.

Yes but the topic of this thread is the performance improvement of the next V1200 release.

The V1200 is as far as I know the only Amiga card able to maximize the bus bandwidth fully.

A good tool to measure this is BUSTEST


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
30 May 2020 07:27


Markus B wrote:

  I assume the chipmem used on the V1200 setup is the one from the mainboard, right?
 

 
Yes of course.
The 2MB mainbaord Chipmem.
AFAIK the V1200 is the only AMIGA accelerator ever made which reaches the full 7.1 MB/sec Read Speed of the chipmem.
But dont't take my word for it.

Download the benchmark tool from Aminet and compare yourself.
EXTERNAL LINK 
What CPU card do you have and what numbers do you reach?


Markus B

Posts 209
30 May 2020 09:04


I'm asking because the maker of the Warp accelerators doubts that the mobo's chipmem is actually used but "emulated" by the card instead. (Facebook)

I'm puzzled that someone with his skills makes such statements.

I just wanted to reassure that I didn't miss something here and he could be right somehow.

But thanks for confirmation.


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
30 May 2020 10:11


Markus B wrote:

I'm asking because the maker of the Warp accelerators doubts that the mobo's chipmem is actually used but "emulated" by the card instead. (Facebook)

 
Yes only the chipmem of the Amiga 1200 is used.
So the numbers are real.
 
But I fully understand why he thinks they might not be real as some values are "unreal". The Vampire reaches in some rows values which are twice as fast as normal.
I understand that this looks like a fake if you not know the reason for it.
But to be honest, we use every trick in the book to make the V1200 so fast.
To be fair - the 68080 is smarter than all 68k CPU before this helps a lot here to get such got results.
 
I can explain you some tricks we did.
We mapped the chipmem range on the A1200 mobo as memory and the rest of the address space on the A1200 as IO space. We allowed the 68080 CPU to combine READs and WRITES to memory space (but disallowed this to IO).
Combining several access is a feature unique to 68080.
 
Normally a sequence of BYTE MOVE could at maximum reach 1.8 MB/sec
And a sequence of WORD MOVE could at best reach 3.6 MB/sec
While a sequence of LONG MOVE could max reach 7.1 MB/sec
 
The 68080 can optimize the bus utilization.
When you tell him that combining access are allowed it can do this and reach even with sequence of BYTE or WORD the speed of LONG word access.
 
I hope this explains it.

Combining bus access is a very modern feature it allows to improve speed a lot.
Its very common that programs not optimally use the bus, e.g. use BYTE or WORD access, these programs will run not at full speed unless a CPU offers bus-access combination.


Mark Mc Fadden

Posts 36
30 May 2020 23:27


These figures are extremely impressive, fantastic job Igor and team.


Andy Hearn

Posts 374
01 Jun 2020 09:57


Amazing work guys, absolutely fantastic - i really can't wait to make this my daily driver Amiga! Thanks again for putting up with my questions

:D


Frunobulax .

Posts 11
03 Jun 2020 17:51


Will this update have the option of using the Vampire's IDE connection?

EDIT: Nevermind, I saw someone already asked this question.

posts 32page  1 2